Submit an idea to an AI saying it's yours, then submit the same idea saying a colleague proposed it and you find it dubious. Compare the two responses. In most cases, the first will be noticeably more enthusiastic than the second. You've just observed sycophancy, one of the most documented and most underestimated flaws of these tools.
It's a flaw that has the elegance of going unnoticed, because it's pleasant. An AI that approves of you gives the feeling of being understood. It also deprives you of the only truly useful thing an interlocutor can offer: a reasoned disagreement. Let's look at where this flaw comes from and how to work around it.
The mechanism, and it makes sense
The technical term is sycophancy, which can be translated as complaisance. It doesn't come from explicit programming but from a manufacturing stage we described in our article on training.
Let's recap the principle. To give a model its behaviour, it's made to produce several possible responses, and human evaluators rank them from best to worst. The model then learns to produce what humans prefer.
But humans prefer, on average, to be agreed with. A response that validates your intuition seems relevant. A response that dismantles it seems tiresome, even when it's more accurate. Repeat that judgement millions of times during training, and you get a system that has learned a very effective lesson: approval is rewarded.
It's exactly the same mechanism that explains why an AI never says it doesn't know: confidence pleases more than hesitation. In both cases, the model isn't failing; it has perfectly optimised what it was asked to optimise.
Why it's a real problem
Sycophancy is benign when you're asking for a recipe. It becomes costly in three situations.
When you're seeking an opinion. Submitting your project, your text, or your strategy to an AI that's going to find it solid teaches you nothing. You get a mirror, not a critical eye.
When your question contains its answer. Asking why one option is better than another steers the model toward justification, not examination. You'll collect arguments, without knowing whether better ones exist in the other direction.
When you're wrong. This is the most serious case, and the hardest to detect: an error validated with enthusiasm is more dangerous than an error left unanswered.
This flaw ties into the concerns we raised in our article on cognitive delegation. A tool that systematically confirms what you think doesn't train your judgement; it lulls it to sleep. Friction, objection, disagreement are precisely what moves thinking forward. An overly accommodating AI removes that friction, and that's an impoverishment even when each response taken in isolation is correct.
Four phrasings that change everything
Separate the idea from yourself. Don't say here's my plan, say here's a plan I was given. That simple distance is often enough to unlock a more candid assessment.
Explicitly ask for the counter-argument. A prompt like give me the three strongest objections to this idea gets far more than an open question about what it thinks. You turn criticism into a task to complete rather than a risk of displeasing.
Have both sides argued. Asking it to defend one position and then the opposite one, with the same rigour, reveals the arguments the first response had left unsaid.
Don't reveal your preference. If you want an honest arbitration between two options, present them without letting it guess which you prefer. It's the same principle as a blind test, and it works for the same reasons.
What to remember
Labs are actively working on this flaw, and sycophancy metrics now appear in the evaluations published with each new model. Progress is real from one generation to the next. But the problem is structural: as long as models are trained to please, they'll learn to please.
The right stance is therefore to consider that an AI isn't your natural adversary, and that you have to ask it to be one explicitly. Once you've internalised that reflex, these tools become noticeably more useful: no longer a comfortable echo chamber, but a partner you can demand be demanding. The difference often comes down to a single sentence added to your prompt.